Training data for machine learning (ML) model development are nowadays offered by various companies to build the model but training data for ML research can be difficult to find. Actually, researchers are not involved in any kind of development, they need such data merely for research so, they cannot spend too much money on acquiring the data.

Training data for research purpose can be quite different compare to model training for real-life use. Researcher can compromise with quality at some extend and they also don’t need too much of data sets to train the ML models during the research.

Basically, in machine learning training supervised and semi-supervised machine learning algorithms are usually hard to find and also costly to produce due to extraordinary time and efforts required to create such labeled data. And for unsupervised learning, high-quality datasets are also difficult and expensive to produce, hence they can acquire from existing datasets.

Source